- What is speech?
- Speech applications
- Speech databases
- Speech hierarchy
Today we start the next part of the course in which we go deeper into individual topics in multilingual nlp and we will talk about morphosyntax and we talk about morphosyntax because the same meaning can be grammaticalized in different ways in different languages for example some languages will grammaticalize the meaning of future as word others as morpheme and others as grammatical structure and today we will talk about morphology morphological processing morphological analysis and inflection generation and in the next class we will talk about syntax we’ll start with a reminder from the beginning of the course when we discussed linguistic typology we have discussed earlier the definition of a word what is a word is a slippery topic and we have agreed on the following definition that words are the building blocks of phrases and sentences that are members in a syntactic category so they have a part of speech that words tend to be able to move around relatively to each other and be composed into sentences so until now what i mentioned is a syntactic definition and words are linguistic units that usually have main stress sometimes as secondary stress this is a phonological definition and they can correspond to a unit of mean this is a semantic definition so all this together is a definition of a word which works rather robustly although there are still some problems for example related to idioms and multi-word expressions such as an expression by and large now words themselves are not atoms they have internal structure and morphology is the study of the formation and internal structure of words as you can see in this drawing a fundamental concept in morphology is that of morphine morphine is a minimal pairing of form and meaning by form we mean the sequence of sounds like on the bottom of this drawing 3 and it can be a sequence of letters for example spelling of the word 3 and this is a mapping between the here the phonetic string 3 and the meaning or content of the word and this mapping is called is a morpheme and we say that morpheme is a minimal pairing because we require that morpheme is a pairing that cannot be reduced to smaller subunits so for example the plural form of the word three trees will contain two more themes three and s plural marker okay so words are made of morphemes and what i highlighted in a sentence on the top or in the example of rich morphology of english on the bottom is what you can see is the concatenations of [Music] morphemes and there are three morphemes morphemes that occur independently for example the word establish and there are bound morphemes those that are attached to other words so in the word this establishment established is a free morpheme but their prefix this and suffix meant are bound more things words can undergo morphological processes or formal operations what are these morphological processes these are for example concatenation it is the most common morphological process if we concatenate the stem which is the core of the word with the affixes which are purely grammatical morphemes that cannot occur in isolation this process is called a fixation so in the example this establishment the stem is established and the there are prefix and suffix these and meant and the concatenation of if you concatenate prefix and then stem and the suffix both from two sides this process is called circumfixation so these are concatenative processes uh there are also non-concatenative processes uh for example when we add an infix in there so this when we add a morpheme in the middle of a string and another common type of morphological process is compounding when rather than concatenating bound morphemes we concatenate three morphemes we concatenate several stems like in the word dishwasher or in the word skyscraper so you can see that there are two types of morphological processes and these morphological several types of morphological processes more than two and these morphological processes can change the meaning of the word like for example with compounding or with [Music] the the modification of establishing to disestablish or they can change the grammatical function for example cats turning is a plural form of the word cat but the meaning has not changed here are some examples of interesting languages and their morphologies tagalog language is spoken in philippines it illustrates a great many form of morphological processes including prefixations fixation in fixation and reduplication so here the singular form is formed with prefix ma and then the plural form is formed with both prefixing prefixation and prefixing reduplication the infix boo and you can read more examples of interesting examples in the book that i reference below by laurie levin and david mortensen this is still a draft but a very nice draft human languages for artificial intelligence here are some other interesting examples arabic is a semitic language that uses templatic morphology in these languages root consists of two or four consonants and then morphology is uh represented by vowels and consonants which are inserted in between the root consonants so the root and the pattern of specific vowels each function like separate morphemes uh so they are combined like a beads in the stream and the uh on the bottom we can see chinese so although chinese has poor inflectional morphology it has many compound words and so there are compounding processes you can see examples below which i cannot read okay morphological functions so talking about morphological processes uh such as concatenation so compounding only tells us part of a story our next question in what are the functions of these operations the two broadest functional categories of morphology are derivation and inflection derivational morphemes examples are on the top are used to create new words they change meaning they can result in a new part of speech but of course not always for example the same example establish and disestablish and then another modification is this establishment then when we move from the verb to noun the second type of morphological function is uh and the important class of morphemes are inflectional morphemes they use to mark grammatical distinctions like cat and the suffix s will turn it into a cat but it does not change the meaning of the word so how do we formally discuss morphological properties of a language and learn about languages morphology in linguistics so in linguistics we introduce a interlinear gloss text igt interlinear gloss text is a standard way to explain the structure and properties of the language via examples it is described in detail in a document that i link on the top of the slide and i recommend you read this document if the topic of morphology is interesting to you and on the bottom i paste the simplest example of igt interlinear gloss text from the same document so you can see that there are three lines the top line corresponds to the original source language or object language which is in our case indonesian the third line is a translation of the first line into a meta language which is typically in linguistic papers this is english this is a lingua franca the second line is the gloss and it is between the first and the third one this is a line this is why it’s called interlinear every token in the second line is written in a meta language in english but it is not an english sentence this is an alignment with talking with tokens in their object language so you can see that the word they in jakarta now are aligned with the first line all tokens are aligned and this is important because this gloss contains information more complicated examples of this gloss contain information about morphosyntax of the object language which is indonesian and in this example we can see from this even simple example of igt we can see the difference between english and indonesian and in particular how english uses copula verb r b while indonesian doesn’t and here is another more complicated example of the interlinear gloss text between the morphologically rich russian this is our object language the top line and the bottom line is that translation of the sentence into english which is a meta language and in between you can see two lines representing a little bit different information this is optional both variants are possible this these are interlinear gloss text so the word the the fourth token in the russian sentence [Music] is split into three more themes corresponding to stem past tense and plural and you can see this specification in the igt lines two and three so this would be a stem go and then past tense and the number plural and the word bas of tobusan its fifth token in russian is split into two more themes corresponding to stem and suffix marking the case instrumental case so you can see from this linguistic description and alignment between the two sentences and the morphological analysis of individual talking in the top sentence how the morphologies of russian and english are different and you can see that morphology of russian is significantly more complex than the morphology of english in this slide which i will not read to you but you can use it later as a reference i list the types of morphological categories across parts of speech all languages possess the same set of about 25 categories number gender tense in if each of which have several functions so there are roughly 25 categories and roughly 100 functions for example a category for nouns a category [Music] number can of the object can be singular plural dual across different languages this corresponds to number another example of a category is noun class which uh noun classes are linked to grammatical gender in indo-european languages but in other languages this can be arbitrary set of categories and all nouns must belong to a category so an interesting example of noun classes uh is in swahili which is an african bantu language uh there are several classes i don’t remember i think it’s seven there is a class corresponding to anime animate nouns denoting people animals insects birds fish there is a different class corresponding to trees plants body parts and another chorus class corresponding to foods and so on and there is a separate class for singular animated nouns versus plural animate now so these are semantic classes and the different morphological rules different prefixes and suffixes are applied to different classes so languages differ in how they express these categories some use like seems like chinese vietnamese these are morphologically what we call poor languages some use freestanding grammatical morphemes such as pronouns prepositions so which means that they express the same categories but through syntax and some languages use affixes prefixes suffixes and this means that they express these categories through morphology so this is what it means when i said in the beginning of the class that different languages grammaticalize the this relationships and the cutting function categories and functions differently through words or through syntax or through morphology and this is why we have different types of languages which we categorize into morphological typology so there is also a disagreement on the typology but i present uh an analysis from david morton sense and the lori levine’s book and traditionally languages have been divided into four times types based on how morphologically rich they are languages like chinese vietnamese and a little less english are called isolating or analytic languages these languages have relatively little morphology and very little inflectional morphology with the exception of compounding like i showed in the example of chinese so most words consist of a single frame or theme languages that similarly to isolating languages have relatively few three more themes per word but unlike isolating languages have many prefixes and suffixes are called fusional or flexional languages examples include greek russian german there is a special subclass of fusional languages which are called templatic languages as i showed in previous example arabic and hebrew is also like this the next category is agglutinative or agglutinating languages these languages often have many more themes within a world examples include finnish turkish and swahili and these languages have very long words and finally the most morphologically rich languages are polysynthetic which can include the whole sentences whole propositions uh they can concatenate them into a single topic so analytic languages have lower morpheme to word ratio and higher restrictions of word order in a sentence languages with richer morphology synthetic languages typically have higher morpheme to word ratio in more free word order there are tight token curves that also can visualize how [Music] properties of morphologically rich languages so we can see that the richer the morphology of a language the higher is type to token ratio in a language and consequently the higher is the sparsity of tokens in the corpus here for example [Music] this paper by 2012 they took a multi-parallel corpus consisting of six languages with varying degrees of morphological complexity so these are the same sentences this is a multi-parallel corpus but written in different languages in english chinese german arabic turkish and korean and you can see that as the morphological complexity of the language increases the number of types in the corpus increases resulting in a steeper curve the most morphologically rich in this drawing is innoctitude which is a polysynthetic language it concatenates many morphemes together into single words and on top of this it has it has complex morphophonological uh processes between morphemes so uh in for inuktitut at 1 million tokens it it it has approximately 225 000 types compared to english which is for around 1 million tokens has only 30 000 types okay so due to morphological complexity and the diversity of languages there are several challenges in processing morphologically rich languages first is related to high lexical sparsity due to the variability of morphological forms this leads to sparse statistics many forms of words appear only once or few times even in very large corpus and there are many plausible word forms that do not appear at all these are called out of vocabulary words or ovs and these out of vocabulary or rare words that appear only once so twice lead to errors in systems such as machine translation speech recognition dialogue question answering and so on so on so this is the first challenge of processing morphologically rich languages the corpora no matter how large they are they are very sparse but on top of this morphologically rich languages are often resource poor so the problem of sparsity is much more acute specifically for these languages an additional challenge of morphologically rich languages is their complex word agreement this leads to agreement errors in language generation especially when we need to model agreement between words that appear far from each other in a sentence just because as i mentioned in some of the in the few slides before that the morphologically rich sentences are languages also have more free word order okay so on top of this high variability there is also a mismatch between morphologies of variability between morphologists across languages and language families and this is so mapping between morphologies of different languages is difficult this leads to difficulties in transfer learning when we want to transfer training corpora from language one language to another uh this can lead to problems this can also lead problems in translation errors for example when we translate from languages that don’t have the concept of definiteness like chinese or russian and so chinese and russian don’t have definite indefinite articles and when we would like to translate from chinese or russian into english that have definitely indefinite articles so these translations tend to contain errors and another problem is that differences in mapping across morphologies of different languages can lead to exacerbation or amplification of social biases for example when we translate from languages like turkish or hungarian that do not mark gender of third person pronouns like he or she and we translate from turkish or hungarian into english so it was shown in previous work that for example translation of lower status occupation would be translated with the female pronouns whereas the translations of higher status occupations would be translated with male pronouns so a turkish sentence that is written basically they are a nurse would be translated as she is a nurse and they are a scientist would be translated as he is a scientist and this just amplifies biases in english training data so to summarize main challenges for processing morphologically rich languages are high sparsity different errors in generation for example caused by uh in for the need to enforce agreement between words and the differences in morphological properties across languages that are different to map across languages and because of these challenges we do morphological processing and incorporate morphological knowledge into models explicitly so what are the types of morphological processing first morphological analysis and we can talk about morphological parsing or morphological segmentation the second is related to generation and the task can be inflection generation or full paradigm completion i will show later an example and the third less common task would be in acquisition of morphology although we have resources for hundreds of languages where we know their morphology uh thousands of languages we still don’t have enough knowledge about their mythological properties so the task of acquisition of morphology would focus on learning languages morphologies automatically so the most common task is morphological analysis and also this it is called morphological parsing our input is a word and our output is a word stem and features expressed by other morphemes for example given in input cats morphological analyzer’s output will be cat plus n which corresponds to noun plus pl which corresponds to the plural form so morphological parsing is the process of determining what are those morphemes given an input word and as you can see in these examples uh this information morphological information comes from three major sources first from lexicon so we need to know the stem second from morpho tactics we need to know morphosyntactic rules and third is from spelling or pronunciation rules because we need to know uh for example um the world cities uh we need to know that as a stem city which ends with a y in a plural form will be changed into cities with eye having this information about lexicon about morphotag tactic rules and about spelling rules we can build a morphological analyzer classical approaches to morphological analysis used finite state transducers and they achieved the high accuracies high results good results so uh in these approaches we have two tapes uh lexical and surface as in this example in the drawing on the top these drawings are from the girovsky martin textbook so and we build a finite state transducer so either one so it can be inverted so either lexical form can be in in surface form can be an input and the output can be stem plus morphological analysis or stem plus morphological analysis can be an input and the output would be a surface form so we will build a transducer so that given an input word cats it would read the word the input letter c and output c lead and read a output a read t output t then read the plus n and would output an empty string epsilon and then it would read plus pl plural and would output s so the output would be cats and it’s coming from lexical interpretation to the surface interpretation so on the bottom we have a simplified fst for morphological parsing cats goes from q0 to q1 from state q0 to state q1 this is a regular noun then q1 outputs an epsilon the for for the noun plus n and then we have we can go from q4 to output s a plural form and we have the end of string in q7 so we can build such an fst finance state transducer as a recognizer and it would say is it a legal string if we if we accept it and or if it’s not a legal string if it’s not a word illegal word in a language in this case the fst would reject it and we can build this fst as a translator in which we input a string of characters from the surface form and output the string of morphemes okay so in 2016 there were a surge of papers that started to replace fin state transducers with current neural networks so this is one of the many examples of papers that i present here by katarina khan and her collaborators ryan cotterell and henrik schutzen and they showed how to implement the same task of morphological analysis using recurrent neural networks and i present specifically this paper because it’s interesting for two reasons first they show how to do canonical segmentation rather than surface level segmentation of words so canonical segmentation so you can see in this example for the word achievability uh how would for the input achievability how would an output surface segmentation look like and how would output canonical segmentation look like so for a surface segmentation we will just segment the input sequence into substrings in canonical segmentation we don’t only segment the sequences into substrings but also replace substrings with the canonical form this canonical segmentation has several representational advantages [Music] because it shows more themes that are not obfuscated by morphology so it reduces the sparsity more significantly but it is also more challenging for example it would be difficult to implement with finite state transducers uh because it doesn’t only segment uh words into substrings but also reverses orthographic changes so how did they do it they first used the standard character based bi-directional uh gru encoder with attention and then they have a unidirectional gre decoder so this is a very standard sequence to sequence architecture uh run end with attention and then a novel component they introduced was a neural re-ranking for segments to identify canonical segments so in particular when they kind of produce the segmentation of a word they first output [Music] a surface level segmentation but then they did re-ranking and specifically looked for segments that are more frequent when they occur as an independent word is in a lexicon and for this they incorporated simple lexical information into word embeddings that marked an additional that added an additional dimension to the word embedding a binary feature that said whether this word is appears as an independent word in the lexicon or not so this is a standard sequence to sequence character-based model plus a re-ranker on top of it to prioritize words that appear more frequently as independent words in the lexicon so in case in this case the in the segmentation of a word achievability the word so the second token able would appear more frequently than the token abil so how do we evaluate morphological analysis there are several evaluation measures that are described in the paper in the previous slide and in several other papers first is error rate defined as 1 minus the proportion of guesses that are completely correct so proportion of 1 minus proportion of outputs of the model that are completely correct second is the added distance so when the first first is a hard measure it’s uh the correct or incorrect output eleven stain distance would measure the similarity between the gas and the gold standard form and finally morpheme f1 score calculates the precision and recall of correctly generated morphemes where the training corpora usually come from in this paper and other papers on morphological analysis a most common resource to work with is unimorph it annotates hundreds of languages using a unified schema it includes currently annotations for 110 languages and i put the link in the slide okay so this was about morphological analysis uh the second type of prominent type of direction of research is focusing on morphological generation in particular inflection generation there are two tasks one is uh to generate a single inflection so our input would be a lemma and grammatical features for example [Music] this is an example in german cow which is a calf and the grammatical features such as case nominative number plural infraction generator will output calvary and the second task which is a generalization of inflection generation task introduced relatively recently in single phone competitions that given a lemma as an input we would need to fill a table and this table this means we need to general generate all possible inflections of a word for all morphological categories so an example where inflection generation either inflection generation or paradigm completion is useful is for reducing sparsity of the corpus for example in machine translation machine translation suffers from data sparsity as i mentioned before when translated thing morphologically rich languages since every surface form is considered every token is considered as an independent entity and to alleviate the problem of sparsity we can translate into lemmas in the target language and then apply inflection generation as a post-processing step so these tasks morphological analysis morphological inflection generation paradigm completion they are widely explored in particular in the sigmar phone competitions so sigma phone is a special interest group on computational morphology and for phonology and they organize workshop and workshops and share tasks every language and i put here some interesting summaries of workshop tasks that you can look into and there is a lot of research so i really gave a high level overview morphological inflection generation in general the base model is a usually a simple bile stem with attention so some recurrent sequence to sequence model uh and then while the base model is simple there are several interesting challenges specific to morphology an interesting challenge for example is how to enforce monotonic alignment between input and output because we unlike in machine translation morphology we don’t need to do a lot of reordering to output to the output sequence uh second interesting challenge is how to better learn longer distance relationships between characters to model variation in inflection due to spelling and phonological rules in a language so some of the papers focusing on shared tasks specifically focus on for example adding a crf layer on top of the decoder to better model relationships between characters in the outputs and if you want to know more about the field you can start with the sigma phone 2020 share task summary which provides an overview and the historical development of the task and the links to interesting papers and to read a particularly interesting paper about morphological generation let’s read this paper that i put in this slide so it particularly focuses on low resource settings so morphological inflection has been pretty thoroughly studied in monolingual high resource setting but in this paper adonis and graham focus on um low resource perspective and uh on cross-lingual transfer across languages and your task for the discussion is just to read this paper and provide your insights for example you can provide the critique about an individual component of the model so the model in the paper consists of several different components so you can talk about an individual component or individual experiment or experimental setup and in particular you can propose ideas for follow-up work or how a specific modeling decision could be adapted to a language of your choice for example a language that you speak and that’s it thank you and see you tomorrow
Outline
Here is a lesson outline on words and morphology, based on the sources:
- What is a word?
- Words are the building blocks of phrases and sentences. They have a part of speech and can move around relative to each other in a sentence.
- Words usually have main stress (phonological definition) and correspond to a unit of meaning (semantic definition).
- It can be difficult to define “word” robustly due to idioms and multi-word expressions.
- Orthographic definition: strings separated by white spaces.
- Prosodic definition: words have one main stress.
- Syntactic definition: words are the syntactic building blocks of sentences.
- Semantic definition: words are units that describe a single idea or semantic concept.
- Structural Subfields of Linguistics
- Phonetics: The study of the sounds of human language
- Phonology: The study of sound systems in human languages
- Morphology: The study of the formation and internal structure of words
- Syntax: The study of the formation and internal structure of sentences
- Semantics: The study of the meaning of sentences
- Pragmatics: The study of the way sentences with their semantic meanings are used for particular communicative goals
- Parts of Speech
- Open classes: nouns, verbs, adjectives, adverbs
- Closed classes: prepositions, determiners, pronouns, conjunctions, auxiliary verbs
- Part-of-speech tags with descriptions and examples
- Morphology
- Words have internal structure, and morphology is the study of the formation and internal structure of words.
- A morpheme is a minimal pairing of form and meaning. It is a pairing that cannot be reduced to a smaller subunit. For example, the plural form of the word ‘three’—‘trees’—contains two morphemes: ‘three’ and ‘s’ (the plural marker).
- Words are made of morphemes.
- Free morphemes can occur independently (e.g., “establish”), while bound morphemes are attached to other words (e.g., “dis-” and “-ment” in “disestablishment”).
- Morphological Processes
- Morphological processes are formal operations that words undergo.
- Concatenation is a common morphological process where a stem (the core of the word) is combined with affixes (grammatical morphemes).
- Affixation involves concatenating a stem with affixes.
- Circumfixation involves concatenating a prefix, stem, and suffix.
- Adding an infix involves adding a morpheme in the middle of a string.
- Compounding is when free morphemes are concatenated (e.g., “dishwasher,” “skyscraper”).
- Morphological processes can change the meaning of a word or its grammatical function.
- Morphological Functions
- Derivation: Derivational morphemes are used to create new words, changing meaning and potentially the part of speech (e.g., “establish” to “disestablish” to “disestablishment”).
- Inflection: Inflectional morphemes mark grammatical distinctions without changing the word’s core meaning (e.g., “cat” to “cats”).
- Interlinear Gloss Text (IGT)
- IGT is a standard way to explain the structure and properties of a language via examples.
- It typically consists of three lines: the original source language, a gloss (token-by-token alignment with tokens in the object language) in a meta-language (typically English), and a translation into the meta-language.
- IGT can show differences in how languages express grammatical information.
- Morphological Categories
- Languages possess roughly 25 categories, each with several functions.
- Examples include number (singular, plural, dual) and noun class.
- Languages differ in how they express these categories (through words, syntax, or morphology).
- Morphological Typology
- Languages can be categorized based on their morphological richness.
- Isolating/analytic languages have little morphology (e.g., Chinese, Vietnamese).
- Fusional/flexional languages have many prefixes and suffixes (e.g., Greek, Russian, German). Templatic languages such as Arabic and Hebrew are a special subclass of fusional languages.
- Agglutinative/agglutinating languages have many morphemes per word (e.g., Finnish, Turkish, Swahili).
- Polysynthetic languages concatenate entire propositions into a single word (e.g., Inuit).
- Analytic languages have a lower morpheme-to-word ratio and stricter word order, while synthetic languages have a higher morpheme-to-word ratio and more flexible word order.
- Languages can be categorized based on their morphological richness.
- Challenges in Processing Morphologically Rich Languages
- High lexical sparsity due to the variability of morphological forms.
- Complex word agreement, leading to errors in language generation.
- Mismatch in morphologies across languages, complicating transfer learning and translation.
- Morphological Processing Types
- Morphological Analysis/Parsing/Segmentation: Identifying the word stem and features expressed by morphemes (e.g., “cats” -> “cat” + noun + plural).
- Morphological Generation: Inflection generation or full paradigm completion, generating inflections of a word based on its lemma and grammatical features.
- Acquisition of Morphology: Automatically learning languages’ morphologies.
- Morphological Analysis
- Involves determining the morphemes in an input word.
- Information comes from a lexicon, morphotactic rules, and spelling/pronunciation rules.
- Classical approaches use finite state transducers.
- Morphological Generation
- Generating a single inflection or completing a full paradigm (generating all possible inflections of a word).
- Useful for reducing corpus sparsity in machine translation.
- Related NLP Problems
- Tokenization
- Lemmatization
- Text normalization
- Spelling correction/grammatical error correction
- Syntactic tagging and morphological analysis
- Evaluation of text generation or machine translation (on the word level)
- Resources and Tools
- UniMorph: A project annotating morphological data in a universal schema.
- SIGMORPHON: Special interest group on computational morphology and phonology.
- Finite state morphology tools: Xfst, FOMA
- Unsupervised methods such as Morfessor
- Stemming
- Byte-pair-encoding (BPE)
Papers
Here is a list of papers and resources covered in the lesson outline, based on the sources:
- Laurie Levin and David Mortensen’s book: Human Languages for Artificial Intelligence.
- Interlinear Gloss Text (IGT) documentation: A document describing IGT in detail.
- Paper by Katarina Kann et al. (2016): This paper implements morphological analysis using recurrent neural networks, focusing on canonical segmentation.
- UniMorph: A resource annotating morphological data in a unified schema for hundreds of languages.
- SIGMORPHON workshops and shared tasks: Workshops and shared tasks focusing on computational morphology and phonology.
- Paper by Adonis and Graham: A paper focusing on morphological generation in low-resource settings and cross-lingual transfer across languages.
- Paper by At 2012: A paper that uses a multi-parallel corpus consisting of six languages with varying degrees of morphological complexity.
- Girovsky and Martin textbook: Used in classical approaches to morphological analysis with finite state transducers.
- Sylak-Glassman (2016): A paper specifying the schema of UniMorph.
Citation
@online{bochman2022,
author = {Bochman, Oren},
title = {Morphology},
date = {2022-03-01},
url = {https://orenbochman.github.io/notes-nlp/notes/cs11-737/cs11-737-w18-morphology/},
langid = {en}
}